Every education company says technology should serve teaching rather than replace it. It is an easy sentence to write and a difficult one to hold to, because the moment you mean it, you start turning features off. A principle only becomes real at the point where it costs you something — a slower answer, an extra click for a busy teacher, a demo that looks less magical than a competitor's. Here are five of those points, and what each one costs.
The tutor will not simply hand over the answer
A student can ask our tutor a homework question and will not get a finished homework answer back. It explains the concept, shows the method, and returns a question that checks whether the student followed it. It will not write an essay for a student, and for an assigned problem it walks through the reasoning rather than stating the final number. This is the least commercially attractive decision in the product. A tutor that instantly produces the answer feels more powerful in the first thirty seconds and is worth less by the end of the term, because the thing being practised is the reasoning, not the retrieval.
Nothing the AI produces becomes a decision on its own
Every teacher-facing tool in the platform generates a draft, and a draft is where it stops. Lesson plans, worksheets, question papers, slide decks, marks suggested from a scanned homework page — all of it lands in an editor with a person's name on the save button. There is no path in the software where AI output reaches a class, or a report card, without a teacher passing through it. Written into the agreement we sign with each school, and into the product, is that the following stay with people:
- grading and the final mark on any piece of work
- assessment of a student's progress
- curriculum delivery — what is taught, in what order, at what depth
- educational supervision
- any disciplinary decision
The cost is real: a teacher who wanted a finished exam paper in one click has to read and edit it instead. We think that is the correct amount of friction. The teacher knows which children in that room will be defeated by question four, and no draft knows that.
A question outside the syllabus gets a deflection, not a guess
Answers are grounded in the CDC textbooks and the school's own approved material — the tutor searches that corpus at the moment a question is asked and cites the book, chapter and page it drew from. In Exam Mode, the wall is absolute: if the retrieved material does not support an answer, the question never reaches the model at all. The student sees a single line — that this is beyond the current coursework, and to bring it to their teacher in class. A general-purpose assistant would happily have produced four confident paragraphs there. During revision week, a plausible answer from outside the syllabus is not a bonus; it is a student memorising something they will not be examined on, with no teacher in the loop to catch it.
Inclusive Education has no AI in it at all
This one was a deliberate absence rather than a constraint. Our inclusive-education module is a set of structured observation forms a teacher fills in about a child on a given day, a full revision history of every edit, and a school-level rollup so a coordinator can see when several teachers have independently noticed the same thing. There is no model anywhere in it. It documents what an adult observed; it does not screen, label or classify any child, and a count of flagged observations is a prompt to look closer and involve someone qualified — never a finding about the student. Noticing how a child learns is skilled human work performed by someone who knows them, and automating it would have made the output faster and worth less than nothing.
The parts that motivate can be switched off
Streaks, badges and leaderboards work well for some students and quietly corrode others — the same counter that makes one child try harder makes another afraid to open the app after missing a day. So every one of them can be turned off, by the student or their guardian, without turning off the learning underneath. It costs us the engagement numbers that make a product look healthy in a monthly report. A child who stopped practising because the streak became something to lose was never really engaged anyway.
And we say out loud that it can be wrong
AI systems produce answers that are sometimes incomplete, outdated or wrong, and a school buying one deserves to hear that from the vendor rather than discover it in a staff meeting. It is stated plainly in our terms and repeated in training: review what matters educationally, never rest an important academic decision on AI output alone, and treat everything it produces as a draft by a capable assistant who has never met your students.
None of these choices make for a better demonstration. Each of them makes for a better classroom, which is the only measure we can defend to a head teacher a year from now. Technology second does not mean technology small — it means the software gets the second-largest vote in the room, every time.